Why Identification Requires More Than Conditional Means
A conditional average treatment effect (CATE) describes how the average causal effect varies across values of pretreatment covariates. For a covariate profile \(X=x\), it is tempting to subtract the observed mean outcome among treated individuals from the observed mean among untreated individuals. But those groups may differ in ways that also affect outcomes. The subtraction is a causal effect only under assumptions that connect treatment assignment, potential outcomes, and the observed data.
This tutorial derives that connection for a binary treatment and explains what each condition contributes. The result is a population identification formula: it expresses a causal estimand in terms of population quantities that can, in principle, be learned from observed data. In a finite sample, the corresponding conditional means are estimated, not known exactly.
The definition uses potential outcomes, only one of which is observed for each individual. Identification asks whether \(\tau(x)\) can nevertheless be expressed using the distribution of observed variables. For each covariate profile of interest, the standard conditions are consistency, conditional exchangeability, and treatment positivity.
The Three Identification Conditions
These conditions address distinct problems. Consistency connects an observed outcome to a potential outcome. Exchangeability allows the potential outcomes in a treatment group to represent those in the full covariate stratum. Positivity ensures that each group is present there, so both observed conditional outcome means are defined from data on that stratum.
Identification Formula and Justification
Under the three conditions, each potential-outcome mean conditional on \(X=x\) equals an observed conditional outcome mean in the treatment group that received the corresponding intervention. For either \(a=0\) or \(a=1\), the argument is:
The first equality holds because \(Y=Y(a)\) when \(A=a\). The second follows because conditional exchangeability makes the conditional distribution of \(Y(a)\) the same among those with \(A=a\) and among all individuals with \(X=x\). Positivity ensures the conditional mean on the left can be formed for both treatment values. Subtracting the equality for \(a=0\) from the equality for \(a=1\) gives the identification result:
The formula does not say that any observed difference between treated and untreated individuals is automatically causal. Its causal interpretation depends on all three conditions. Nor does identification guarantee that an estimate from a particular sample is precise: a stratum can have positive treatment probabilities but very few observations in one arm.
Worked Example: Identifying a Population Risk Difference
Consider a hypothetical population with binary outcome \(Y\), where \(Y=1\) denotes an adverse event. At covariate profile \(X=x\), suppose the population conditional outcome means are \(E[Y\mid A=1,X=x]=0.30\) and \(E[Y\mid A=0,X=x]=0.18\). Suppose consistency and conditional exchangeability hold, and \(P(A=1\mid X=x)=0.40\), so both treatment probabilities are positive.
Because the outcome is binary, each conditional mean is the corresponding event risk. The identification formula gives
In this hypothetical population, the identified CATE is an increase of \(0.12\), or 12 percentage points, in the adverse-event risk under treatment for individuals with \(X=x\). The calculation is valid as a causal interpretation because the assumptions were stipulated to hold. The same subtraction without exchangeability, consistency, or positivity would not by itself identify a causal effect.
From Identification to a Sample Estimate
In practice, population conditional means are unknown. For a binary outcome, the sample proportion with \(Y=1\) in each treatment arm and covariate stratum estimates the corresponding population conditional mean. Their difference is a plug-in estimate of the identified CATE, not the exact population CATE. Identification assumptions do not remove sampling variability or make finite-sample proportions equal to population risks.
Worked Example: Estimating a Conditional Risk Difference From Counts
In a hypothetical sample restricted to \(X=x\), 32 of 80 treated individuals experience the event, while 24 of 120 untreated individuals experience it. Assume for interpretation that the three identification conditions are appropriate for this stratum. The sample includes individuals in both arms, as required for estimating both conditional risks.
The estimated conditional event risks are
The estimated conditional risk difference is therefore
Thus, the sample estimates a 20-percentage-point increase in event risk under treatment for this covariate profile. The population CATE is identified, under the assumptions, by the difference in the population conditional means. These finite counts estimate that difference; they do not establish that the population CATE is exactly \(0.20\). Sampling uncertainty would need to be considered for an inferential conclusion.
The same distinction applies when covariates define several strata. Each stratum’s sample risk difference estimates its own population CATE if the identification conditions hold there. Similarity or difference among those estimates is not, by itself, proof that the population effects are identical or heterogeneous.
What Goes Wrong When a Condition Fails?
The assumptions are not interchangeable. Two hypothetical examples show why observed conditional means alone can be misleading or insufficient.
Worked Example: Unmeasured Confounding Within a Covariate Stratum
Let \(U\) be an unmeasured binary risk factor within a fixed stratum \(X=x\). Suppose, for every individual, \(Y(1)=Y(0)=U\): treatment has no effect for anyone. In a hypothetical population, let \(P(A=1\mid X=x)=0.50\), and suppose \(P(U=1\mid A=1,X=x)=0.80\), while \(P(U=1\mid A=0,X=x)=0.20\).
Consistency holds, and positivity holds because each treatment arm has probability \(0.50\). But conditional exchangeability given \(X\) fails: treatment groups differ in \(U\), which determines both potential outcomes. The observed conditional risks are \(0.80\) among treated individuals and \(0.20\) among untreated individuals, so their difference is \(0.60\). Yet the causal effect is
The observed 60-percentage-point difference is an association generated by the imbalance in \(U\), not a treatment effect. Measuring and appropriately conditioning on relevant confounders can support exchangeability, but the condition cannot be verified just by checking whether the observed groups look similar on measured variables.
Worked Example: A Missing Treatment Arm at a Covariate Profile
Suppose that in a hypothetical population every individual with \(X=x\) receives treatment, so \(P(A=1\mid X=x)=1\). The observed treated outcome mean at that profile is \(E[Y\mid A=1,X=x]=0.40\), but there are no untreated individuals there. Consistency can link the observed treated outcomes to \(Y(1)\), but there is no observed conditional mean for \(Y(0)\).
For example, two possible populations could have the same observed treated outcomes and the same treatment assignment pattern, but different unobserved untreated potential-outcome means. If \(E[Y(0)\mid X=x]=0.10\), the CATE is \(0.40-0.10=0.30\). If instead \(E[Y(0)\mid X=x]=0.70\), it is \(0.40-0.70=-0.30\). Both possibilities are compatible with the observed treated mean. Without positivity, the observed data do not determine which CATE is correct.
Common Mistake: Treating an Estimate as the Identified Population Effect
A common error is to calculate a difference of sample proportions and report that the assumptions make this exact difference the population causal effect. The assumptions justify a population identity; they do not eliminate the gap between sample estimates and population quantities. The correct sequence is to state the population identification formula, calculate the sample estimates of its two conditional means, and describe their difference as an estimate of the population CATE.
A second error is to treat positivity as a mere technicality. If one treatment arm is absent at a covariate profile, the missing arm’s conditional outcome mean cannot be recovered by subtracting observed means. Extending conclusions to that profile requires additional assumptions or a different target population; it does not follow from the identification formula.
Check Your Understanding
Answer each question by distinguishing the assumptions, the population identity, and the sample quantities.
- State the conditional exchangeability assumption using \(Y(0)\), \(Y(1)\), \(A\), and \(X\). What makes it relevant to the causal interpretation of a conditional mean difference?
- In a stratum, the population conditional outcome means are \(0.25\) for treated individuals and \(0.10\) for untreated individuals. If all three identification conditions hold, what is the CATE and how should it be interpreted for a binary adverse-event outcome?
- A sample gives estimated event risks of \(18/90=0.20\) among treated individuals and \(24/120=0.20\) among untreated individuals. What is the estimated risk difference, and why does this not establish that the population CATE is exactly zero?
- Which identification condition fails if all individuals with a particular \(X=x\) receive treatment? What information about the CATE is then missing from the observed data?
- Suppose treatment and control groups have different distributions of an unmeasured cause of the outcome within \(X=x\). Which condition is threatened, and why can the observed conditional risk difference then differ from the CATE?